Developing ChIMES Machine Learned Interatomic Potentials for Cesium Lead Halide Perovskites Through Hierarchical Learning

Interest in perovskite materials for optoelectronic applications has surged due to their impressive photovoltaic performance demonstrated nearly a decade ago. [1] Initially, research focused on organic-inorganic hybrid perovskites [2], but their relatively poor stability has led to a shift towards the more stable cesium based inorganic lead halide perovskites. [3] While inorganic perovskites offer high conversion efficiency compared to silicon-based materials, they suffer from rapid degradation in the presence of water and relatively poor thermal stability. [4] Recent studies suggest that modifying synthesis protocols, such as solution processing, can enhance material stability, yet the underlying nucleation mechanisms remain poorly understood. [5, 6] Although solution-processing is theoretically highly tunable, the multitude of chemical compounds involved in perovskite synthesis complicates the optimization of governing conditions.<br/>Simulations can offer mechanistic insights to address these challenges. In this work, we present a ChIMES physics-informed, machine-learned model development approach to create interatomic potentials for characterizing cesium-lead halide perovskite. [7–10] Our ChIMES model features a unique hierarchical structure that allows parameter fitting to be broken down into manageable and reusable "bricks." [11] In the present work, this modular approach starts with binary element pairs and systematically builds to a comprehensive cesium-lead halide perovskite model. We demonstrate that training our ChIMES model using randomly perturbed structures can improve the efficiency of model generation compared to similar models generated using complete dynamics runs.<br/>Our approach can enable the efficient development of accurate machine-learned interatomic potentials to probe the effects of halide segregation on the structural properties of the perovskite framework and characterize diffusion pathways that facilitate halide partitioning. The flexible ChIMES framework also allows for the extension of the model to include solvent interactions.<br/><br/><i>References</i><br/><i>[1] Akihiro Kojima et al. </i><i>J. Am. Chem. Soc. 131.17 (2009). </i><br/><i>[2] Mingchao Wang et al. </i><i>J. Mater. Chem. A 8 (34 2020). </i><br/><i>[3] Rachel E. Beal et al. </i><i>J. Phys. Chem. Lett. 7.5 (2016). </i><br/><i>[4] Loredana Protesescu et al. </i><i>Nano Lett.</i> <i>15.6 (2015). </i><br/><i>[5] Adnan Younis et al. Adv. </i><i>Mater. 33.23 (2021).</i><br/><i>[6] Dan Li et al. </i><i>ACS Omega 5.29 (2020). </i><br/><i>[7] Rebecca Lindsey et al. J. Chem. Theory Comput. 13.12 (Dec. 2017).<br/>[8] Rebecca Lindsey. ChIMES Calculator. URL: https://github.com/rk-lindsey/chimes_calculator.<br/>[9] Rebecca Lindsey. ChIMES Generator. URL: https://github.com/rk-lindsey/chimes_lsq.<br/>[10] Rebecca Lindsey. ChIMES Active Learning Driver. URL: https://github.com/rk-lindsey/al_driver.<br/>[11] Rebecca Lindsey et al.</i> <i>ChemRxiv.</i> <i>(2024).</i>